Human review is often asserted but rarely evidenced.
Many AI governance approaches assume human review provides accountability. BAT repeatedly encounters workflows where review is described, but the evidence of meaningful review is unclear.
BAT
Research Programme
Testing accountability against operational reality.
BAT is an active research methodology for examining whether AI-supported reliance can be reviewed, reconstructed, and defended after the moment of use.
The programme is deliberately open to challenge. Its purpose is not to confirm a finished framework, but to test where accountability controls hold, fail, or need to be redesigned.
Reliance means a person, team, or organisation uses an AI-supported output as part of a decision, action, communication, recommendation, prioritisation, or other consequential activity.
Research Question
What controls actually make AI-supported reliance defensible?
What BAT Has Learned So Far
Early findings from testing BAT against workflow reality.
Review capacity appears to be an accountability control in its own right.
Assigning a human reviewer is not the same as designing a review process. As workflow volume increases, accountability may depend as much on review capacity as on review authority.
Accountability often becomes unclear at the point of reliance.
Policies, governance structures, and oversight arrangements may exist, but responsibility can become ambiguous when an AI-supported output is actually used to make a decision, take an action, or communicate with others.
Not every consequential workflow requires the same evidence.
BAT's early testing suggests accountability responses may need to be proportional to context, rather than relying on a single standard of documentation for all AI-supported activity.
Accountability may be evidenced without recording every prompt.
Early BAT development suggests that reconstructability may depend more on ownership, review activity, rationale, escalation, and evidence objects than on exhaustive prompt logging.
The presence of a human may not be the strongest accountability control.
BAT began with many of the assumptions common in AI governance discussions. Ongoing testing has raised the possibility that, in some workflows, review design, challenge mechanisms, escalation pathways, and evidence quality may contribute more to accountability than the simple presence of a human reviewer. This remains an active hypothesis under investigation.
Current Hypotheses
Claims being tested, not assumptions being sold.
Human review may not always be the strongest accountability control.
Accountability can be evidenced without recording every prompt.
Review capacity must be operationally designed, not merely assigned.
Aggregate consequence may require distinct accountability treatment.
Seeking
The programme needs cases that make the method work harder.
Challenge cases where ordinary governance language does not explain what happened.
Pilot workflows where AI-supported reliance is already entering operational practice.
Research collaborators working on accountability, assurance, audit, or responsible AI.
Critical feedback on where BAT is unclear, incomplete, or too easy on itself.
Participate
Help test whether BAT survives real workflow conditions.
The BAT Challenge collects real and representative workflow examples to test the methodology against operational detail. Submissions should avoid confidential, personal, or identifying information.
Open the BAT Challenge